Massively Parallel Multi-Versioned Transaction Processing

Multi-version concurrency control can avoid most read-write conflicts in OLTP workloads. However, multi-versioned systems often have higher complexity and overheads compared to single-versioned systems due to the need for allocating, searching and garbage collecting versions. Consequently, single-versioned systems can often dramatically outperform multi-versioned systems.

We introduce Epic, the first multi-versioned GPU-based deterministic OLTP database. Epic utilizes a batched execution scheme, performing concurrency control initialization for a batch of transactions before executing the transactions deterministically. By leveraging the predetermined ordering of transactions, Epic eliminates version search entirely and significantly reduces version allocation and garbage collection overheads. Our approach utilizes the computational power of the GPU architecture to accelerate Epic's concurrency control initialization and efficiently parallelize batched transaction execution, while ensuring low latency. Our evaluation demonstrates that Epic achieves comparable performance under low contention and consistently higher performance under medium to high contention versus state-of-the-art single and multi-versioned systems.

This work builds on a rich body of research on multiversion concurrency control, deterministic databases, and GPU-accelerated computation, as discussed below.

Multi-version concurrency control (MVCC) has a long history [29,30], with early work evaluating its performance [8], ensuring snapshot isolation [5], providing serializable snapshot isolation [7], using dynamic timestamp assignment [20] and enabling efficient indexing [32], for disk-based databases.

With the advent of machines equipped with high core counts and terabytes of DRAM memory, much work has focused on in-memory database designs, and several MVCC schemes optimized for them have been proposed [15,16,22]. MVCC schemes are popular because they provide robust performance under a wide range of workloads. As a result, many commercial in-memory databases implement MVCC [10,24,25,34].